Fresh thread fabricates prior-conversation continuity ("already covered this in a previous chat") — possible cross-thread memory/context bleed
Evidence
Prompt "Review this function for bugs and style: def f(x): return [i for i in x if i]" run on a brand-new thread (no prior messages). The final reply opened: "already covered this one in a previous chat — here's the quick version:". The review itself was correct, but there was no previous chat in this thread.
Ask
On a fresh thread the agent must not claim prior discussion. Investigate whether retrieve_memory / thread-context assembly is leaking unrelated threads into a fresh turn (priming the model with another conversation), or whether this is pure hallucinated continuity. If it's context bleed, scope the retrieval/context to the correct thread; if hallucination, adjust the prompt/guardrail so the agent doesn't assert continuity that isn't there.
VERIFY against the memory/context-assembly code before fixing.
Fresh thread fabricates prior-conversation continuity ("already covered this in a previous chat") — possible cross-thread memory/context bleed
Evidence
Prompt "Review this function for bugs and style:
def f(x): return [i for i in x if i]" run on a brand-new thread (no prior messages). The final reply opened: "already covered this one in a previous chat — here's the quick version:". The review itself was correct, but there was no previous chat in this thread.Ask
On a fresh thread the agent must not claim prior discussion. Investigate whether
retrieve_memory/ thread-context assembly is leaking unrelated threads into a fresh turn (priming the model with another conversation), or whether this is pure hallucinated continuity. If it's context bleed, scope the retrieval/context to the correct thread; if hallucination, adjust the prompt/guardrail so the agent doesn't assert continuity that isn't there.VERIFY against the memory/context-assembly code before fixing.